Homeostasis

Designing complex agentic systems in your head becomes impossible very quickly. As soon as you combine several layers of memory, specialized agents, conditional logic, and data flows moving between services, the architecture becomes too abstract to visualize. You end up writing code for weeks just to test a hypothesis, without knowing whether the real mechanism will hold up in production.

Today, the tools make the situation even more fragmented: one AI for dialogue, another to generate images, third-party services for data, orchestrators to stitch everything back together. Each block has its own API, its own pace, its own constraints. You have to synchronize states, maintain overall coherence, monitor heterogeneous architectures… and even then, it remains very hard to observe a complete flow in a simple way, from user input to internal processing. What is missing is a unified environment where you can visually assemble autonomous components, see data moving in real time, and validate an agentic architecture end to end before committing to a heavy implementation.

Homeostasis is that missing layer: a complete environment for designing, observing, and running complex agentic architectures. The canvas interface lets you assemble autonomous nodes, track their activity, then switch without interruption to a headless mode where those same nodes respond to external HTTP requests. The same system is used to imagine, test, adjust… then deploy, without rewriting the logic in another framework.

At the core of Homeostasis is a form of digital cryptobiosis: each node exists in the graph but consumes nothing until it is called. No polling. No looping process. A node activates on an event, data received, interaction, API request, processes, broadcasts what it needs to broadcast, then goes back to sleep. Communication follows a structured peer-to-peer logic: no central coordinator, each node knows only its neighbors and passes messages from one point to the next. This topology makes it possible to build highly responsive architectures while keeping them light and naturally scalable.

Extensibility is native: a new node is defined in four standardized files (JSON structure, PHP processor, communication API, JS interface). Once described, these elements are analyzed and assembled automatically inside the environment; the node appears in the library and can be instantiated like any native component. The system supports several language model providers (OpenAI, Anthropic, Mistral, Venice…), with the ability to switch from one model to another by changing a single parameter.

Technically, Homeostasis fits inside a minimalist architecture: vanilla PHP, SQLite, and JSON, with no required external dependency, deployable on simple shared hosting. This restraint, about 13 MB for the complete environment, is not an aesthetic choice. It is the direct result of the founding principle: run only what is necessary, when it is necessary. The result is agentic architectures that would not be practical to code by hand in the usual way, where autonomous agents collaborate without a central orchestrator, and where global complexity emerges from simple, controlled interactions.

Homeostasis is built visually. Each component is autonomous, each connection is direct, each interaction can be observed in real time. The following demos show how the system turns abstract architectural concepts into concrete working mechanisms.

Homeostasis continues to evolve with the pace of progress in artificial intelligence. The chatbot engine will gain native file handling: reading, splitting, storage, document recall, and image analysis with OCR. The goal is for agents to handle reports, notes, or screenshots as naturally as they handle text. Narrative memory will also be refined to include dynamic evolution of the system prompt, so that an agent’s personality and behavior can gradually adjust to the context built over the course of the exchange.

External connectivity will expand toward everyday tools: integration with Google Calendar, Gmail, Drive, and exploration of compatibility with the MCP protocol to make connections with other intelligent systems easier. The idea is to let agents interact directly with the digital ecosystem, without going through a jungle of specific APIs. Video generation will be integrated natively, so nodes can produce complete multimedia content when the use case requires it.

The tools system will become fully modular: instead of burying technical constraints in the system prompt, an activation mechanism will allow capabilities to be plugged in or unplugged as needed. The prompt will remain dedicated to the agent’s role and personality, while tool-related instructions will be managed separately. This separation simplifies configuration, clarifies context, and makes it possible to quickly test different combinations of providers and models depending on the task.

Finally, nodes that are isolated today will gradually receive standardized connectors so they can fit into larger architectures. Homeostasis is not a fixed product. It is a living environment, meant to grow with the technical capabilities of models… and with the real uses made by those who put it into production.